README.md

Weight Agnostic Neural Networks

Weight Agnostic Networks: network topologies evolved to work with a variety of shared weights. Adapted from the prettyNEAT package. This repository represents a snapshot of the code used to produce the results in the original WANN paper. The goal was to create minimal rather than general code for clearer understanding. If you are interested in a more extendable and general implementation for your own experiments, we recommend using the WANN fork of the prettyNEAT package.

Dependencies

Core algorithm tested with:

• Python 3.5.3

• NumPy 1.15.2 (pip install numpy)

• mpi4py 3.0.1 (pip install mpi4py)

• OpenAI Gym 0.9.6 (pip install gym -- installation details here)

Domains tested with:

• Cart-pole Swing-up (included, but requires OpenAI gym)

• Bipedal Walker: Box2d (see OpenAI gym installation)

• Quadruped (Ant) Walker: PyBullet 1.6.3 (pip install pybullet)

• MNIST: Mnist utilities 0.2.2 (pip install mnist)

• VAE Racer:

  • Tensorflow 1.8 (pip install tensorflow==1.8.0)
  • Pretrained VAE (in wannRelease -- copy to root to use, e.g: cp -r ../vae .)

Training Weight Agnostic Neural Networks

To get started and see that everything is set up you can test the swing up domain:

python wann_train.py

which is the same as the default hyperparameter:

python wann_train.py -p p/laptop_swing.json -n 8

Where 8 is the number of workers you have on your computer, and p/laptop_swing.json contains hyperparameters.

Evaluation of the population is embarrassingly parallel so every extra core you have will really speed things up. Here is an example training curve on cart-pole with a population of 64 on an 8 core laptop:

Where Fitness is the mean reward earned over all trials and weight values. Median Fitness is the fitness of the median performing member of the population, Max Fitness is the fitness of the best performing member of the population, Top Fitness is the best performing member ever found. Peak Fitness is the mean reward earned by the best performing member with its best performing weight value. To reproduce this graph see this jupyter notebook.

The full list of hyperparameters and their meaning is explained in hypkey.txt

Testing and Tuning Weight Agnostic Neural Networks

To view or test a WANN:

python wann_test.py -p p/swingup.json -i champions/swing.out --nReps 3 --view True

WANNs are saved as 2D numpy arrays and can be retested, train, and viewed using the WANNTool provided.

---

Amira

General

To run on windows change line 236 in wann_train.py to subprocess.check_call(["mpiexec","-np", str(n), sys.executable] +['-u']+ sys.argv, env=env) instead of subprocess.check_call(["mpirun","-np", str(n), sys.executable] +['-u']+ sys.argv, env=env)

Frep

python wann_train.py -p p/frep.json -n 2

Rover training

How to add new training "myName"

1. in ./p add myName.json file and change task name
2. in ./domain
   1. add myName.py
      1. change class to myNameEnv
   2. in make_env.py
      1. add else if statement
   3. in config.py
      1. add games['myName'] entry and change layer sizes and activation function types

how to show the results/plot graph